docs(roadmap): final v2.x is v2.0.2; split v3.0.0 (integrity) and v3.1.0 (features) - #403

Syncs ROADMAP.md with the milestone re-sequencing already applied on GitHub:

• v2.0.2 (renamed from v2.3.0): final planned v2.x patch. Format v7, no API change. The merged perf: point query on a bound entity scans the entity's whole version history — O(retract/reassert count), not O(live facts) #323/perf: attribute scan is unbounded to end of AEVT index for attributes ending in a 0xBF/0x7F byte #381/perf: checkpoint() is O(graph size) and flat in dirty bytes — it rebuilds all four covering indexes every call #315 mitigations, plus durability: fsync the parent directory after creating the .graph/WAL and after deleting the WAL #389, ci(bench): regression alerts have been muted since #301; fix alerting and noisy thresholds #393, ci: add MSRV check, public-API semver check, and dependency policy check #395, docs: align stability and support messaging with what the project actually does #397, process: keep data-integrity bugs visible until the fix ships in a release #399, docs: define support tiers for language bindings and platforms #400. Documents the bug: same-transaction multi-valued facts read back as ONE value, differing between EAVT and AEVT; a batched retract can leave retracted values visible #371 known issue and its workaround.
• v3.0.0: format v8 and data integrity. bug: same-transaction multi-valued facts read back as ONE value, differing between EAVT and AEVT; a batched retract can leave retracted values visible #371/bug: EavtKey and AevtKey missing value bytes causes silent data loss when batching multi-valued facts #287, storage: add per-page checksums to fact pages and B+tree pages (file format v8) #388, save() is not crash-atomic: new fact pages overwrite old index pages before header commit #374, Add index integrity check (verify) and public rebuild-indexes-from-fact-pages #373, perf: compute net-assert on v8 index keys before resolving facts (O(live) point queries) #379, fuzz: btree_page target never reaches B+tree node decoding #375, plus the release-gate tests and docs from tracker Tracker: production readiness for high-stakes deployments #383 (test: SIGKILL crash test must verify committed data, not only that the file opens #384–fuzz: keep the corpus between nightly runs and add an operation-sequence target #387, test: drive the fault-injection backend through the full save/WAL/recovery paths #390–test: long-running soak test at scale (millions of facts, many checkpoints, reopen cycles) #392, perf: benchmark queries at 100K–10M facts and publish a performance envelope #394, ci(coverage): raise storage/WAL coverage gates and add branch coverage gates #396, docs: durability, recovery and operations guide #398).
• v3.1.0 (new): additive features. Rule persistence: deeper design consideration #241, post-1.0: build query profiler #185, No working result-limiting/pagination mechanism for plain (non-recursive) :find queries #306, Feature request: server-side :limit clause for Datalog queries #310, feat(query): add set membership predicates #316, post-1.0: add lag/lead window functions #182, post-1.0: add sliding row frames for window functions #183, post-1.0: expose PreparedQuery over UniFFI #181, Expose UDF registration in every binding #180, feat(ffi): expose OpenOptions over UniFFI — the close-time checkpoint cannot be suppressed from Python #322, research: temporal graph traversals (temporal network analysis) #273.
• The v2.1.0 and v2.2.0 milestones were deleted after being emptied. The v2.0.0 milestone was closed.

Docs only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BedF9ZmqVZLkvDwn91fuKy